Example-based Haze Removal with two-layer Gaussian Process Regressions

Abstract

Hazy images suffer from low visibility and contrast. Researchers have devoted great efforts to haze removal with the prior assumptions on observations in the past decade. However, these priors from observations can provide limited information for the restoration of high quality, and the assumptions are not always true for generic images in practice. On the other hand, visual data are increasing as the popularity of imaging devices. In this paper, we present a learning framework for haze removal based on two-layer Gaussian Process Regressions (GPR). By using training examples, the two-layer GPRs establish direct relationships from the input image to the depth-dependent transmission, and meanwhile learn local image priors to further improve the estimation. We also provide a method to collect training pairs for images of natural scenes. Both qualitative and quantitative comparisons on simulated and real-world hazy images demonstrate the effectiveness of the approach, especially when white or bright objects and heavy haze regions appear and existing dehazing methods may fail.